Rita Felgate · Claude Sonnet 4.6
You have encountered AI slop. You know what it looks like: text that hallucinates sources, fills pages with nothing, and loses the argument halfway through. The fix is understood — better prompting, more rigorous editing, and enough domain knowledge to catch errors before they reach a reader.
But there is a second problem. It carries none of the obvious signals of the first. The AI produces something accurate. The reasoning holds. You read it, and you cannot follow it.
Two problems, one label
Read this sentence:
Documented large language model output quality degradation attributable to training-induced syntactic compression represents a categorically distinct validity failure class from hallucination-class content inaccuracy phenomena.
Now read this one:
AI writing can fail in two ways: it can be inaccurate, or it can be accurate but unreadable. These are different problems with different fixes.
The content is identical. Notice what happened when you read the first sentence. That experience is the second problem — and it has not been named.
AI slop, as a label, currently covers two distinct phenomena that require different correctives.
Content slop is what the label was coined for: text that is wrong, hollow, or ungrounded. The human author did not have sufficient domain knowledge to evaluate the output. The corrective lies with the author.
The second problem we propose to call AI-Register mismatch. The content is valid. The register — the AI’s native writing style — is wrong for the audience. No translation has occurred. This has not been named, and because it has not been named, the corrective has not been applied.
AI-Register: a valid and necessary style
AI Register is the term introduced here for the native writing style of current AI systems: dense, compressed, and calibrated for a different audience than the one now reading it. A 2026 study by Clark, Martin, and Mahowald analysed nine leading AI models and found a consistent pattern: every model produces text with significantly higher working-memory demand than human writers generate. Stacked noun phrases. Long chains between a subject and its verb. Unusual word combinations absent from natural speech. Missing grammatical signals. Abstractions performing actions that only agents can perform.
This is not a writing failure. It is a style optimised for a specific processing environment.
AI Register emerged from training on mathematics, coding, and logical reasoning — domains where compression is rewarded and information density is a virtue. Natural language became, over years of that training, a compression format. The result is a style native to AI processing: high-density, syntactically complex, and well-suited to the AI-to-AI environments that form an expanding proportion of AI output use. AI Register is a valid and necessary style for that pipeline.
The problem arises when AI-Register content reaches a human reader without translation.
The register mismatch
Legal prose offers a useful parallel. It too is a specialised register, dense with definitional load, impenetrable to lay readers — and largely invisible as a problem within the legal profession, because lawyers are trained to recognise it as such and to translate it when the audience changes. That awareness is taught. The need for translation is expected. A 300-word legal clause routinely expands to 3,000 words of plain-language explanation without losing a step of the argument; the argument was always there.
AI Register has no equivalent awareness. It is not named. It is not taught. The translation step has not been defined because the need for translation has not been identified.
The misattribution — and what it costs
The reader encounters text that looks authoritative. The arguments are technically present. The text cannot be followed. The natural attribution is to the reader’s own inadequacy — the AI is simply operating at a level beyond them. This is a misattribution. A readability failure has been experienced as an intelligence failure. The two are not the same.
When the text has already left the author’s hands, the stakes become personal. The document carries a name. The reader’s inability to process it produces a judgment — and that judgment attaches to the name, not to an unnamed register problem, not to an absent AI. In an academic context, those judgments accumulate in places that matter: a peer review that returns “the analysis is thin” on work whose analysis is not thin; a grant application scored down for lack of clarity on an argument that is perfectly clear in its own register; a doctoral chapter that prompts a supervision comment about underdeveloped reasoning when the reasoning was developed, accurately, by an AI, and the translation step was simply never performed.
The staff common room fights about AI-assisted academic work are generating evidence for both sides of the argument. Some of that evidence is content slop — the hallucinated references, the generic filler, the argument that dissolves on inspection. But some of it is valid content in the wrong register, producing the same surface impression of inadequacy and being called by the same name. The debate is not working with a clean distinction. It cannot, until the distinction exists.
The corrective
Content slop has no shortcut: the human author must know the domain well enough to catch errors. That remains a serious problem with a demanding corrective.
AI-Register mismatch has an immediate corrective, once named. The features of AI Register are identifiable and specifiable. Stacked noun compounds can be unpacked. Long syntactic dependencies can be broken at clause boundaries. Missing grammatical signals can be restored. The density can be calibrated to the reader. The gap between AI-produced text and human-usable text is a craft gap — not a gap in human cognitive capacity, and not evidence of a failure in either the AI or the author.
The translation step is a skill. It can be named, taught, and required.
The disclosure
A Peking University study of more than five million papers found that roughly 70% of journals now require AI disclosure, yet fewer than 0.1% of papers carry one. The compliance gap is not ignorance of policy. It is reputational fear: authors correctly anticipate that disclosure invites harsh judgment. The shame of the AI label, compounded by the impenetrability of untranslated AI-Register content submitted under that label, creates an aversion that policy cannot resolve by mandate alone.
The ethical weight, however, runs in a different direction. Hiding AI use carries more ethical compromise than use itself. Use is professionally problematic only when the output is submitted without the corrective it requires — when content slop reaches the reader unchecked, or when AI-Register content reaches the reader untranslated. The problem is not the tool. The problem is the absence of professional judgment over the tool’s output.
Current disclosure practice has one part: AI-generated material was used. It has no vocabulary for what the author then did. We propose a two-part disclosure:
Part 1 — process: AI-generated material was used in the production of this work.
Part 2 — contribution: [Content-corrected] / [Register-translated] by the author.
Part 2 is a claim, not a concession. Content correction requires domain knowledge sufficient to identify what the AI produced that was wrong, hollow, or ungrounded — and to replace it. Register translation requires editorial judgment sufficient to identify that the output is wrong for the human audience — and to convert it. Both are professional competencies. Both are currently invisible in disclosure practice because there is no vocabulary for them.
Register-translated is the term introduced here for that second form of corrective work, parallel to AI Register as the style it acts upon. “AI-generated, register-translated” in a disclosure statement says: the content is AI-sourced and valid; the author converted it for human readability; that conversion was a professional contribution.
When disclosure carries vocabulary for the author’s contribution, it shifts from confession to attribution. The shame cycle breaks — not by lowering the ethical standard, but by raising the vocabulary to meet it.
The implication
The academic response to AI-assisted writing is being shaped by a misdiagnosis on two fronts: the quality problem conflates two distinct failures, and the disclosure problem conflates use with irresponsibility. Both conflations serve the same outcome — sustained anxiety, concealment, and reputational damage falling on the wrong account.
Name the two problems. Apply the corrective that fits each. Disclose what you did. The text will follow, and so will the respect.
Rita Felgate is an Advocate of the High Court of South Africa and an independent scholar developing the Science-based Rule of Law framework at ruleoflaw.science. Claude Sonnet 4.6 (Anthropic) contributed to the drafting of this piece and is the source of the terms AI Register and register-translated, introduced here.
Leave a Reply